Forest Fire Detection Method Based On Machine Learning

The occurrence of forest fire has a great impact on the economy,the ecological environment and so on.Forest fires break out hundreds of thousands of times every year throughout the world,which seriously endangers the earth’s ecological resources and the survival and development of mankind.Therefore,prevention and surveillance of forest fires have been focusing by researchers around the world.The difficulty of video image type forest fire detection is that the current technical means are mainly traditional image detection methods,that is,manually extracting features and then training the classifier for recognition.In view of the complexity of the forest environment,the accuracy of the algorithm will be lower,the robustness is not strong and the applicability is poor.In view of the existing problems,this thesis has conducted in-depth research,the main contributions are as follows:(1)Complete the forest fire database by shooting and collecting videos/images online.These include positive and negative samples,as well as normalization of images,production of sample tags,etc.(2)A new suspicious region segmentation algorithm is proposed for the particularity of forest scenes.The method firstly uses the three-frame difference method to roughly extract the suspicious area,and then further separates the flame area and the smoke area based on the RGB and HSV mixed color space model and the dark channel theory,and finally sends it to the classifier for classification.Experiments show that this method can segment the flame area and the smoke area more accurately than other methods.(3)Aiming at the problems of poor recognition,misdetection and serious missed detection in traditional machine learning,this paper proposes a method combining deep learning with traditional machine learning classifier.The method firstly trains a deep learning model offline.When training the classifier,the image is input into the trained deep learning model for feature extraction,and then the feature is trained as the input of the classifier.The experimental results show that the classification method proposed in this paper is much better than the traditional LeNet-5 and traditional SVM,which not only solves the problem of low recognition accuracy of traditional classifiers,but also overcomes the problem that deep learning models can not effectively deal with small Problem with sample data.(4)Aiming at the problem of incomplete feature extraction from deep learning model,this paper proposes a method to fuse deep and shallow features of flame and smoke.The deep feature refers to the feature extracted by the neural network model.Because the sample is less,the model can not be fully trained,resulting in incomplete extraction features.The shallow features of artificially extracted colors and textures also contain rich information.Get a more complete feature.The experimental results show that the method improves the accuracy of forest fire identification.Finally,the work of the full text is summarized.